about the pretrain model "90k" and "90k_bn"

[irfanICMLL]

I have evalute the miou of these two pre train model.
The results are as follows:
90k: 78.56%
90k_bn: 70.34%
Where does the pretrain model "90k_bn" come from?
Can I fune-tuning the "90k" model?
Thanks a lot for your implementation, and looking forword for your reply~~:)

[hellochick]

Hey @irfanICMLL .
The pretrained model "90k_bn" comes from the source code of ICNet:
https://github.com/hszhao/ICNet/tree/master/evaluation/model
You can fine-tune on it, but I think it cannot be much better.

The reason is, this model have been already pruned by the original author. So, if you focus on the number of kernels, you will find all the numbers of kernels are half of those in PSPNet.

So the best approach is, fine-tuning on the model before pruning, and then do model compression.
Hope my answer can help you. Btw, I keep training ADE20k dataset on it, and get result in the bottom of README.

[irfanICMLL]

So the decrease of the accuracy is caused by the purning?
And if I want to fune-tuning the "90k", I need to write another script? What is the loss function of the ICNET. I found that the ICNET only has one branch for output, while the ICNET_BN has three and the loss is in a combined form. The loss function described in the artical is also a combined form. I feel really confused....

[hellochick]

@irfanICMLL , I don't think the decrease of the accuracy is caused by the pruning, I still can't find the answer to the strange accuracy. Maybe the model is different to each other.
You can directly modified train.py, I have create the loss formula in it. You can take a look on it.